Title: Image deblurring method based on feature fusion SRN

Authors: Junjia Bi; Lingxiao Yang; Jingwen Zhang; Jianjun Zhang

Addresses: School of Electrical Engineering and Automation, Henan Polytechnic University, 2001 Century Avenue, Jiaozuo, 454003, Henan, China ' School of Electrical Engineering and Automation, Henan Polytechnic University, 2001 Century Avenue, Jiaozuo, 454003, Henan, China ' School of Electrical Engineering and Automation, Henan Polytechnic University, 2001 Century Avenue, Jiaozuo, 454003, Henan, China ' School of Electrical Engineering and Automation, Henan Polytechnic University, 2001 Century Avenue, Jiaozuo, 454003, Henan, China

Abstract: This article proposes a SRN algorithm of feature fusion to solve the problem of image motion blur. First, an Attention Residual Module (ARM) is designed to add channel attention between residual units to increase feature extraction capabilities. Second, a feature pyramid structure is constructed to improve the representation ability of the network. Then, a multi-scale coordinate attention feature fusion structure is built to improve the deblurring effect of the model. Finally, optimising the loss function improves the robustness of model to discrete points and increases the stability of the model. The testing is performed on the GOPRO dataset. Our algorithm is the best, with PSNR and SSIM reaching 34.72 dB and 0.97. Tested on the foreign object data set, the PSNR and SSIM of our algorithm have been greatly improved, and compared with other methods, it has a great advantage in detailed texture recovery.

Keywords: motion image deblurring; feature pyramid network; attention residual module; multi-scale fusion; loss optimisation.

DOI: 10.1504/IJCCPS.2023.133728

International Journal of Cybernetics and Cyber-Physical Systems, 2023 Vol.1 No.3, pp.234 - 245

Received: 01 Jun 2022
Accepted: 17 Jun 2022
Published online: 02 Oct 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article